Effects of serial dependency on the agreement between visual and statistical inference.

The authors built fake data sets that looked like single-case graphs.

Each set had different levels of autocorrelation—high, medium, or none.

They asked both people and a computer to judge if the data showed a real change.

Then they counted how often the two answers matched.

When autocorrelation was high, people and the computer disagreed most of the time.

With no autocorrelation, the two methods usually said the same thing.

The paper warns: if your data points drag each other along, visual and statistical calls can clash.

A year later, Annable et al. (1979) showed that even trained visual raters only agree with each other 61 % of the time.

Together the two papers tell a story: autocorrelation hurts visual-stat agreement, and even eyeball-to-eyeball agreement is shaky.

Branch (2019) picks up the thread forty years on, arguing that built-in replication—not p values—keeps behavior analysis honest.

All three papers push the same fix: tighten your method, don’t just trust the graph or the number.

Before you call a phase change real, run a quick autocorrelation check in Excel or R. If the lag-1 r is above 0.3, show the team both visual and statistical views and explain the mismatch. This small step saves you from later replication headaches.